How do you handle selection bias and confounding variables when analyzing historical user interaction data at Spotify?
